The described technology relates generally to communications techniques and particularly to communications between hosts and data store devices.
The speed and capacity of the data store devices, such as disk drives and memories, have increased significantly over the past several years. As a result of their improved performance, these data store devices are being used in many new applications, such as database servers, Web servers, personal video recorders, and digital displays. These applications often require large amounts of data to be communicated between data store accessing devices (“hosts”) and data store devices. (Hosts may include computers, CPUs, or any logic for accessing a data store device.) Moreover, as host speed increases, the speed in communicating between hosts and data store devices can have a significant impact on the overall performance of the application. In particular, even though the speed of hosts and data store devices has increased significantly, the speed of communications between hosts and data store devices has not increased as significantly, especially for communications over long distances (e.g., greater than one meter). Thus, the communications speed presents a bottleneck in many new applications.
Current communications techniques typically communicate between hosts and certain types of data store devices, such as disk drives, using a bus with many parallel lines or using a single serial communications link. The Integrated Disk Electronics (“IDE”) bus and the Small Computer Systems Interface (“SCSI”) bus are examples of bus-based parallel communications techniques. These communications techniques, however, present many problems. Performance of bus-based communications techniques is generally improved by increasing the number of lines in the bus, which may significantly increase the cost of such techniques. In addition, bus-based communications techniques generally provide arbitration so that multiple hosts and data store devices can share the same bus. The use of arbitration can significantly increase the cost of such a bus. The cost of such bus-based communication techniques is further increased because their design needs to address additional problems such as cross-talk and clock skew. In particular, as the communications speed increases, the solution to cross-talk and clock skew become much more complex.
Some serial communications techniques have been developed to address some of the problems of bus-based communications techniques. Current serial communications techniques, however, have problems of their own. Serial AT, Attachment, which is intended to replace IDE, does not scale well and only operates in a half duplex mode. Fibre Channel, currently used to support storage area networks (“SANs”), is very generic and therefore, not optimized for any particular application. In particular, Fibre Channel has a relatively small packet size with a large header. As a result, use of Fibre Channel often results in an unacceptably large overhead. For example, data transmitted to disk drives is typically sent in very large blocks (e.g., 216 bytes). With Fibre Channel, such large blocks need to be divided in many (e.g., 32) packets, which results in a high overhead in the amount of redundant header information and in the redundant processing performed as a packet is routed to its destination. Thus, Fibre Channel may not be appropriate for many applications.
Current memory devices, such as SDRAM and RDRAM, are typically designed to be synchronous with the accessing processing unit. The hosts and the memory devices are synchronous in that they share the same clock signal. These memory devices are typically optimized for access patterns that are both temporally and spatially related. In particular, these memory devices are optimized to read and write arrays (or streams) of data. There is a setup overhead (e.g., 5 clock cycles) when accessing the first word of an array in memory, but access of subsequent words in the array occurs at the synchronized clock rate (e.g., 1 access per clock cycle). Since the access patterns of central processing units and graphics processors are typically temporally and spatially related, they can access such memory devices efficiently.
Existing memory devices that are designed to support access patterns with a high temporal and spatial relationship may not be appropriate for uses having access patterns with a lower spatial relationship. The setup overhead for each access may be too high. There are, indeed, many uses for memory devices with access patterns that are not as spatially or temporally related as those of a central processing unit or a graphics processor. For example, a switch may have a memory device in which packets of data received via an input port are stored before they are transmitted via an output port. Traditionally, switches used crossbars to provide the switching function and FIFOs to provide a buffering function. When a memory device is used on a switch in place of a crossbar, then all the input and output ports need access to the memory device. The accesses by the different ports are, however, not particularly spatially related. Moreover, when the packet size is small (e.g., 53 bytes in the case of an ATM switch), the spatial relationship of accesses by a single port may not be significant. Other uses in which there may not be a significant spatial relationship of accesses include network processors and caches for storage area networks. In such uses, the data is received from disparate sources at disparate times and may not be spatially related.
Many existing memory devices are not particularly suitable for many uses because the memory devices typically allow access by only one accessing device at a time and because the memory devices typically operate at different clock rates than the accessing devices. Because such memory devices can only be accessed by one device at a time, the accessing devices may need to enter a wait state because the memory device is busy or a memory controller may need to have a buffering component. Of course, the use of a wait state may result in unacceptable performance. Also, the addition of a buffering component may increase complexity and cost. In addition, when multiple accessing devices access the same memory device through a single bus (e.g., one writing to the memory device and the other reading from the memory device), then all the devices that access the memory device need to be synchronized with the memory device. Because the accessing devices may have different underlying clock rates, complex and costly logic is needed to support the mapping to the bus clock rate.
Existing communications protocols, such as Fibre Channel, may have an unacceptable overhead for communicating with memory devices. The communications from a host to a memory device may occur in relatively short blocks (e.g., 32 bytes). Each block needs to be transmitted in a separate packet with a relatively large header. In some packets, the header may be larger than the data itself, which can significantly reduce the overall bandwidth and speed of transmission. More generally, communications between devices typically occurs in a synchronous or an asynchronous mode. In a synchronous mode, the transmitting and receiving devices use the same clock signal. The transmitting device can send the clock signal to the receiving device either as a separate signal or as a signal that can be derived from the data signals. When the clock is sent as a separate signal, problems arise resulting from the different delays in the data signals and the clock signal. These delays and resulting problems are increased as the transmission speed and distance are increased. It is very difficult and costly to account for these delays. In addition, the receiving device will have an asynchronous clock boundary. That is, a portion of the receiving device will operate at the clock frequency based on the transmitting device's clock frequency (i.e., the transmitter's clock domain) and another portion will operate at the receiving device's local clock frequency (i.e., the receiver's clock domain). As a result of the asynchronous boundary, the receiving device typically needs to buffer control and data signals sent between the clock domains using elastic buffers, which adds to the complexity and cost of the receiving devices. These elastic buffers require substantial space (e.g., chip area), and when a single chip has multiple communications ports, the design is complicated because each port needs its own elastic buffer. When the clock is derived from the data signal, the problems of the delay are reduced somewhat, but there are still the problems associated with an asynchronous clock boundary.
A plesiosynchronous clocking technique can be used to avoid the need to transmit a separate clock signal or derive the clock signal from the data signal. With plesiosynchronous clocking (also known as “plesiochronous” clocking), the transmitting and receiving devices have clocks with nominally the same clock frequency. If the clock frequencies were exactly the same, then transmitting and receiving devices would be synchronized and the receiving device could accurately identify the transmitted data (in the case of serial transmission). Also, since the receiving device operates only at its local clock frequency, there is no asynchronous clock boundary. In practice, however, clock frequencies are not exactly the same but vary, for example, by 100 ppm. The receiving device can use techniques as described in U.S. Pat. No. 6,229,859, entitled “System and Method for High-Speed, Synchronized Data Communication,” which is hereby incorporated by reference, to account for clock variations. Those techniques use an oversampling of the data by the receiving device to detect edge boundaries of the transmitted data. The receiving device can vary the number of bits of data detected during an interval to compensate for the variations in frequency.
It would be desirable to have a communications architecture that provides high-performance for applications (e.g., data storage-based applications and memory-based applications) at a low cost. Such a communications architecture would allow for communications techniques to be tailored to particular applications.